Sentinel-2 SuperResolution Models from 10m and 20m to 2.5m

This is a collection of SR models trained on Sentinel-2 data. The models are trained to super-resolve 10m and 20m Sentinel-2 bands to 2.5m resolution. The models are trained on the Sen2NAIPv2. We present three models modalities:

*_SR: Non-reference super-resolution model. The model takes the RGBN 10m Sentinel-2 bands as input and super-resolves them to 2.5m resolution.
*_F2: Reference super-resolution model. The model takes the Red Edges and SWIRs (RSWIRs) 20m Sentinel-2 bands as input and super-resolves them to 10m resolution. It needs that users report RGBN 10m bands as reference.
*_F4: Reference super-resolution model. The model takes the RSWIRs 20m Sentinel-2 bands as input and super-resolves them to 2.5m resolution. It needs that users report RGBN 2.5m bands as reference.

The diagram illustrates the inference procedure for generating a complete 2.5-meter Sentinel-2 data cube. It highlights the specific role of each model in the multi-step resolution enhancement process.

Example Usage

Model	Description	Run Link
Run Simple Model	A lightweight SR model optimized for fast runs.
Run Best Model	Our most accurate model for high-quality super-resolution.
Run Anywhere on Earth	A flexible implementation allowing inference on any location worldwide.

From 10m RGBN S2 to 2.5m

Dependencies for the simple model:

pip install torch supers2 mlstac safetensors

Dependencies for the best model:

pip install torch supers2 mlstac safetensors einops timm mamba-ssm

import torch
import mlstac


# Download model
# file="https://huggingface.co/tacofoundation/supers2/resolve/main/Mamba_Medium_SR/mlm.json" ## best model
file="https://huggingface.co/tacofoundation/supers2/resolve/main/CNN_Light_SR/mlm.json"
output_dir="models2/CNN_Light_SR"
mlstac.download(file, output_dir)

# Create a mlstac object
mlstac_object = mlstac.load(output_dir)
device = "cpu" # "cpu"

# Load model
#srmodel = mlstac_object.trainable_model() # for fine-tuning
srmodel = mlstac_object.compiled_model(device=device) # for benchmarking

# Load Demo Data
lr, hr = mlstac_object.example_data()

# Inference
sr = srmodel(lr.to(device))

# Plot
#fig, ax = plt.subplots(1, 3, figsize=(15, 5))
#ax[0].imshow(lr[0, 0:3].permute(1, 2, 0)*3)
#ax[0].set_title("Low Resolution")
#ax[1].imshow(hr[0, 0:3].permute(1, 2, 0)*3)
#ax[1].set_title("High Resolution")
#ax[2].imshow(sr[0, 0:3].permute(1, 2, 0)*3)
#ax[2].set_title("Super Resolution")
#plt.show()


# Fast plot
fig = mlstac_object.display_results()
plt.show()

From 10m and 20m S2 to 2.5m

import mlstac
import matplotlib.pyplot as plt

# Download model
# file="https://huggingface.co/tacofoundation/supers2/resolve/main/Swin_Medium_F4/mlm.json" ## best model
file="https://huggingface.co/tacofoundation/supers2/resolve/main/CNN_Light_F4/mlm.json"
output_dir="models2/CNN_Light_F4"
mlstac.download(file, output_dir)

# Create a mlstac object
mlstac_object = mlstac.load(output_dir)
device = "cpu" # "cpu"

# Load model
#srmodel = mlstac_object.trainable_model() # for fine-tuning
srmodel = mlstac_object.compiled_model(device=device) # for benchmarking

# Load Demo Data
lr = mlstac_object.example_data()

# Inference
sr = srmodel(lr.to(device))

# Fast plot
fig = mlstac_object.display_results()
plt.show()

Model Details

We provide the following models:

Model Name	Modalities	Description	Input Resolution	Output Resolution	Trainable Parameters (M)	Total Mult-Adds (GFLOPs)
`CNN_Light_SR`	`SR`	Super-resolves 10m Sentinel-2 bands to 2.5m resolution	10m	2.5m	0.4	7.8
`CNN_Light_F2`	`F2`	Super-resolves 20m Sentinel-2 bands to 10m resolution	20m	10m	0.4	7.8
`CNN_Light_F4`	`F4`	Super-resolves 20m Sentinel-2 bands to 2.5m resolution	20m	2.5m	0.4	7.8
`Mamba_Medium_SR`	`SR`	Super-resolves 10m Sentinel-2 bands to 2.5m resolution	10m	2.5m	13.8	69.8
`Swin_Large_F2`	`F2`	Super-resolves 20m Sentinel-2 bands to 10m resolution	20m	10m	32.6	71.5
`Swin_Medium_F4`	`F4`	Super-resolves 20m Sentinel-2 bands to 2.5m resolution	20m	2.5m	13.2	35.4